LlamaIndex ExtractBench

Not yet imported

Schema-guided extraction from enterprise documents. Distinct from the Contextual AI benchmark of the same name.

370 documentsApache-2.0version upstream

Listed, not imported — scores come from LlamaIndex's official leaderboard.

View official leaderboard ↗

From the authors

Leaderboard

Sorted by the selected metric, then lower cost.

upstream ↗ @ 741c1b6 · captured 2026-10-20

Showing value f1; higher is better. 49 systems.

The leaderboard CSV does not report per-system sample counts; missing observations and counts are left unspecified.

Value F1 leaderboard, source: upstream
#	system	value f1 ↓	short—	medium—	long—	¢ / page
1	L LlamaExtract Agentic PlusLlamaIndex · extraction API	96.2%	97.0%	94.3%	94.7%	7.50
2	R Reducto Deep ExtractReducto · extraction API	96.1%	96.8%	94.8%	92.8%	6.35
3	P Pulse (Effort)Pulse · extraction API	96.1%	96.8%	94.7%	93.2%	10.50
4	L LlamaExtract AgenticLlamaIndex · extraction API	95.7%	96.1%	94.8%	94.7%	3.12
5	A Claude Code (Sonnet 5.5 Evidence)Anthropic · coding agent	94.7%	96.2%	93.0%	83.7%	5.69
6	O Codex (GPT-6 Sol Evidence)OpenAI · coding agent	94.6%	96.0%	92.4%	87.3%	9.80
7	P Pulse (Non-effort)Pulse · extraction API	94.5%	95.4%	92.5%	93.3%	3.00

challenge score0100Bold marks the best score in each challenge.

Quality against costValue F1 by cost per page. Systems without positive reported cost are omitted.

extraction API
coding agent
commercial vlm / oss

Head to head

First system

Second system

LLlamaExtract Agentic Plus
RReducto Deep Extract

What it measures

LlamaIndex ExtractBench evaluates extracted values and their supporting evidence across short, medium, and long documents. This is distinct from Contextual AI's ExtractBench.